Modeling Decisions for AI 2026 (Vic, Spain)

MDAI 2026 is the 23rd International Conference on Modeling Decisions for Artificial Intelligence, held from 7 to 9 September 2026 at the Universitat de Vic, about 70km north of Barcelona in Catalonia. It is a working research conference rather than a trade event: three days of paper sessions, plenary talks and a fixed number of parallel tracks, with proceedings published by Springer in the LNAI/LNCS series and distributed at the conference itself. MDAI has run annually since 2004 — Barcelona, Tsukuba, Tarragona, Kitakyushu, Tokyo, Umeå, Mallorca, Milan and, most recently, València in 2025 — and is rated CORE B by the Computing Research and Education Association of Australasia.

The conference’s organising idea is decision-making in the broad sense: how models are built, how information is fused, and how the resulting decisions can be made fair, transparent, explainable and free of unnecessary disclosure of sensitive data. That gives it an unusual shape compared with the general-purpose machine learning circuit. Six tracks structure the call: data science and machine learning; data privacy, covering privacy-preserving data mining, privacy-enhancing technologies and statistical disclosure control; aggregation functions, the mathematical machinery for summarising and fusing information; human decision making, where the interest is in the gap between rational models of decision under uncertainty and how people actually decide; graphs and social networks; and information security, with an emphasis on the security of AI-driven systems.

The published programme shows how that plays out in practice. Monday opens with a plenary from Dr Edgar Galván of Maynooth University, “From Pareto to Semantics: Reframing Multi-objective Genetic Programming through Behavioural Search Spaces”, arguing for a shift from optimisation in objective space to search and reasoning in the behavioural space of what programs actually do. Tuesday’s plenary is Dr Sonakshi Garg of South Asian University, Delhi, on “Beyond the Black Box: Hallucinations and Privacy in LLM”, covering hallucinated outputs, data leakage, prompt injection and mitigation techniques for high-stakes deployment. A third plenary on Wednesday is listed as to be announced.

Paper sessions run in two parallel streams. Expect titles such as detecting LLM hallucinations via anomaly detection on latent representations; memory poisoning and secure multi-agent systems; Shapley values and explainability; fairness-aware aggregation in federated learning; bistochastically private release of longitudinal data; privacy metrics for multi-hop payments on the Lightning Network; information gain estimation for decision trees under local differential privacy; and exact LLM unlearning via targeted relearning. Alongside these sit the more mathematical strands the conference is known for — Sugeno integrals, Möbius fields and non-additive set functions, fuzzy contingency tables, graded logic aggregators — plus applied work on skin cancer classification, wind turbine condition monitoring from SCADA data, clinical kidney data and, memorably, understanding individual cows’ behaviour from activity data.

Programme co-chairs are Vicenç Torra (Umeå University), Yasuo Narukawa (Tamagawa University) and Ramon Reig Bolaño (Universitat de Vic), who is also general chair; the advisory board includes Didier Dubois, Lluís Godo, Janusz Kacprzyk, Gabriella Pasi, Pierangela Samarati and Ronald Yager.

Registration fees are €250 for students at the early rate rising to €450 for non-members registering late, with reduced rates for members of the associated societies; the fee covers coffee breaks, lunches, the gala dinner and a copy of the proceedings, and one registration covers up to two published papers. The social programme includes a city tour of Vic on the Monday followed by a welcome dinner in the cloister. Vic is reached most easily via Barcelona, with a direct bus the organisers recommend over the train. Submission deadlines have closed — 12 May for the LNCS track and 20 May for the USB-only track.

Who should go: researchers and doctoral students working on privacy-preserving machine learning, aggregation operators and fuzzy measures, explainability, federated learning or decision modelling — and anyone who wants a small, genuinely technical European venue where you can talk to authors rather than queue behind them. It is not an industry event, and there is no expo floor.